Prime Intellect open-sources a coding agent that can rewrite its own prompts and sub-agents

Prime Intellect open-sourced Prime Agent today, an AI coding agent whose own prompts, skills, memory, and sub-agents are editable from inside the agent's own run. It is a small architectural break from the fixed-schema harnesses most coding agents still ship with.

Two components carry the design. The first, the Recursive Language Model (RLM), treats the model's context as a variable and routes sub-agent calls as functions inside a persistent read-eval-print loop, so the model can program against its own history. The second, Continual Harness, exposes prompts, skills, memory, and sub-agents themselves as create-read-update-delete operations the agent runs from its own trajectory, including across sessions.

Both papers are preprints, not peer-reviewed, and the claims are the vendor's. The GitHub repo and a single curl installer make the whole stack runnable today.

The launch leaves the central question open: does letting an agent rewrite its own scaffolding add capability, or just add a new way to fail? The strongest version of the skeptic's case surfaced quickly in the Hacker News thread, where developers argued that frontier models no longer need a harness layer at all. No third-party benchmark numbers are in the post.